I built non-autoregressive decision models with RL a year ago
Points and comments are a snapshot, not live.
Open-source Laya decision engine achieves 33ms inference, claims 7-8x speed over Jev.
The author compares Laya to TypeSafe's Jev, arguing their earlier 2025 arXiv paper predates Jev's RLCD approach. Laya uses bidirectional encoders (ModernBERT, mmBERT) for non-autoregressive structured decisions: choice, score, boolean. It claims 32.8ms latency vs Jev's 150ms, open-source Apache 2.0 weights, multilingual routing across 100+ languages, and zero API costs. Performance benchmarks show 0.766 accuracy on typed-decisions benchmark and 0.950 on AG News. Limitations include degraded accuracy with >20 options and need for fine-tuning on domain data.
What commenters are saying
Two camps emerged: those viewing Laya as prior art predating Jev (citing the author's March 2025 arXiv paper and open weights), and those noting distinct use cases (Laya focused on sales/conversations; Jev on type-safe structured outputs). Commenters debated whether the architecture was the same or independently discovered. Several agreed LLMs are overkill for structured classification, with one calling Jev 'BERT with more data'. Some praised Laya's openness versus Jev's proprietary approach, while another questioned whether managing GPUs in production is worth the cost savings.
Some commenters shared their own similar open-source projects, and one debater argued general LLMs outperform custom models on robustness despite inefficiency, drawing analogies to typed vs untyped programming languages.